Using Text Similarity to Detect Social Interactions not Captured by Formal Reply Mechanisms

In modeling social interaction online, it is important to understand when people are reacting to each other. Many systems have explicit indicators of replies, such as threading in discussion forums or replies and retweets in Twitter. However, it is likely these explicit indicators capture only part of people's reactions to each other, thus, computational social science approaches that use them to infer relationships or influence are likely to miss the mark. This paper explores the problem of detecting non-explicit responses, presenting a new approach that uses tf-idf similarity between a user's own tweets and recent tweets by people they follow. Based on a month's worth of posting data from 449 ego networks in Twitter, this method demonstrates that it is likely that at least 11% of reactions are not captured by the explicit reply and retweet mechanisms. Further, these uncaptured reactions are not evenly distributed between users: some users, who create replies and retweets without using the official interface mechanisms, are much more responsive to followees than they appear. This suggests that detecting non-explicit responses is an important consideration in mitigating biases and building more accurate models when using these markers to study social interaction and information diffusion.